Extreme value theory for anomaly detection -- the GPD classifier (Q2027086)

As anomaly detection one understands the task of distinguishing between normal and abnormal test data points. The aim of the paper is to introduce a new classifier for anomaly detection. The second section of the article is dedicated to a short presentation of the main contributions in the domain of anomaly detection, pointing out the importance of the extreme value theory, machine learning techniques or neural networks. An introduction in the univariate extreme value theory with main results and a formal description of the task of anomaly detection are contained in the third and fourth sections of the paper. In the fifth section we find a presentation of the extreme value machine algorithm (EVM) introduced by \textit{E. M. Rudd} et al. [``The extreme value machine'', IEEE Trans. Pattern Anal. Mach. Intell. 40, No. 3, 762--768 (2018; \url{doi:10.1109/TPAMI.2017.2707495})]. Limitations of the EVM are shown. Further, in next part of the article, the authors develop two new kernel free algorithms that perform anomaly detection using extreme value theory. The one called GPDC algorithm focuses on the distance between the point to be classified as normal or abnormal and the points from the training data set. The second algorithm is called GEVC algorithm and is based on the approximation by the GEV distribution. See also [\textit{S. Coles}, An introduction to statistical modeling of extreme values. London: Springer (2001; Zbl 0980.62043)]. The new GPDC and GEVC algorithms are compared with the EVM algorithm on simulated and on real data and performances are largely discussed.